blob: cc770ab7835a436d32781905bbbb5c62fba42604 [file] [log] [blame] [edit]
// Copyright 2022 The Tint Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "src/tint/transform/builtin_polyfill.h"
#include <algorithm>
#include <tuple>
#include <unordered_map>
#include <utility>
#include "src/tint/program_builder.h"
#include "src/tint/sem/builtin.h"
#include "src/tint/sem/call.h"
#include "src/tint/sem/type_expression.h"
#include "src/tint/sem/value_conversion.h"
#include "src/tint/switch.h"
#include "src/tint/type/storage_texture.h"
#include "src/tint/type/texture_dimension.h"
#include "src/tint/utils/map.h"
using namespace tint::number_suffixes; // NOLINT
TINT_INSTANTIATE_TYPEINFO(tint::transform::BuiltinPolyfill);
TINT_INSTANTIATE_TYPEINFO(tint::transform::BuiltinPolyfill::Config);
namespace tint::transform {
/// BinaryOpSignature is tuple of a binary op, LHS type and RHS type
using BinaryOpSignature = std::tuple<ast::BinaryOp, const type::Type*, const type::Type*>;
/// PIMPL state for the transform
struct BuiltinPolyfill::State {
/// Constructor
/// @param program the source program
/// @param config the transform config
State(const Program* program, const Config& config) : src(program), cfg(config) {
has_full_ptr_params = false;
for (auto* enable : src->AST().Enables()) {
if (enable->HasExtension(builtin::Extension::kChromiumExperimentalFullPtrParameters)) {
has_full_ptr_params = true;
break;
}
}
}
/// Runs the transform
/// @returns the new program or SkipTransform if the transform is not required
Transform::ApplyResult Run() {
for (auto* node : src->ASTNodes().Objects()) {
Switch(
node, //
[&](const ast::CallExpression* expr) { Call(expr); },
[&](const ast::BinaryExpression* bin_op) {
if (auto* s = src->Sem().Get(bin_op);
!s || s->Stage() == sem::EvaluationStage::kConstant ||
s->Stage() == sem::EvaluationStage::kNotEvaluated) {
return; // Don't polyfill @const expressions
}
switch (bin_op->op) {
case ast::BinaryOp::kShiftLeft:
case ast::BinaryOp::kShiftRight: {
if (cfg.builtins.bitshift_modulo) {
ctx.Replace(bin_op,
[this, bin_op] { return BitshiftModulo(bin_op); });
made_changes = true;
}
break;
}
case ast::BinaryOp::kDivide: {
if (cfg.builtins.int_div_mod) {
auto* lhs_ty = src->TypeOf(bin_op->lhs)->UnwrapRef();
if (lhs_ty->is_integer_scalar_or_vector()) {
ctx.Replace(bin_op,
[this, bin_op] { return IntDivMod(bin_op); });
made_changes = true;
}
}
break;
}
case ast::BinaryOp::kModulo: {
if (cfg.builtins.int_div_mod) {
auto* lhs_ty = src->TypeOf(bin_op->lhs)->UnwrapRef();
if (lhs_ty->is_integer_scalar_or_vector()) {
ctx.Replace(bin_op,
[this, bin_op] { return IntDivMod(bin_op); });
made_changes = true;
}
}
if (cfg.builtins.precise_float_mod) {
auto* lhs_ty = src->TypeOf(bin_op->lhs)->UnwrapRef();
if (lhs_ty->is_float_scalar_or_vector()) {
ctx.Replace(bin_op,
[this, bin_op] { return PreciseFloatMod(bin_op); });
made_changes = true;
}
}
break;
}
default:
break;
}
},
[&](const ast::Expression* expr) {
if (cfg.builtins.bgra8unorm) {
if (auto* ty_expr = src->Sem().Get<sem::TypeExpression>(expr)) {
if (auto* tex = ty_expr->Type()->As<type::StorageTexture>()) {
if (tex->texel_format() == builtin::TexelFormat::kBgra8Unorm) {
ctx.Replace(expr, [this, tex] {
return ctx.dst->Expr(ctx.dst->ty.storage_texture(
tex->dim(), builtin::TexelFormat::kRgba8Unorm,
tex->access()));
});
made_changes = true;
}
}
}
}
});
}
if (!made_changes) {
return SkipTransform;
}
ctx.Clone();
return Program(std::move(b));
}
private:
/// The source program
Program const* const src;
/// The transform config
const Config& cfg;
/// The destination program builder
ProgramBuilder b;
/// The clone context
CloneContext ctx{&b, src};
/// The source clone context
const sem::Info& sem = src->Sem();
/// Polyfill functions for binary operators.
utils::Hashmap<BinaryOpSignature, Symbol, 8> binary_op_polyfills;
/// Polyfill builtins.
utils::Hashmap<const sem::Builtin*, Symbol, 8> builtin_polyfills;
/// Polyfill f32 conversion to i32 or u32 (or vectors of)
utils::Hashmap<const type::Type*, Symbol, 2> f32_conv_polyfills;
// Tracks whether the chromium_experimental_full_ptr_parameters extension has been enabled.
bool has_full_ptr_params = false;
/// True if the transform has made changes (i.e. the program needs cloning)
bool made_changes = false;
////////////////////////////////////////////////////////////////////////////
// Function polyfills
////////////////////////////////////////////////////////////////////////////
/// Builds the polyfill function for the `acosh` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol acosh(const type::Type* ty) {
auto name = b.Symbols().New("tint_acosh");
uint32_t width = WidthOf(ty);
auto V = [&](AFloat value) -> const ast::Expression* {
const ast::Expression* expr = b.Expr(value);
if (width == 1) {
return expr;
}
return b.Call(T(ty), expr);
};
utils::Vector<const ast::Statement*, 4> body;
switch (cfg.builtins.acosh) {
case Level::kFull:
// return log(x + sqrt(x*x - 1));
body.Push(b.Return(
b.Call("log", b.Add("x", b.Call("sqrt", b.Sub(b.Mul("x", "x"), 1_a))))));
break;
case Level::kRangeCheck: {
// return select(acosh(x), 0, x < 1);
body.Push(b.Return(
b.Call("select", b.Call("acosh", "x"), V(0.0_a), b.LessThan("x", V(1.0_a)))));
break;
}
default:
TINT_ICE(Transform, b.Diagnostics())
<< "unhandled polyfill level: " << static_cast<int>(cfg.builtins.acosh);
return {};
}
b.Func(name, utils::Vector{b.Param("x", T(ty))}, T(ty), body);
return name;
}
/// Builds the polyfill function for the `asinh` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol asinh(const type::Type* ty) {
auto name = b.Symbols().New("tint_sinh");
// return log(x + sqrt(x*x + 1));
b.Func(name, utils::Vector{b.Param("x", T(ty))}, T(ty),
utils::Vector{
b.Return(b.Call("log", b.Add("x", b.Call("sqrt", b.Add(b.Mul("x", "x"), 1_a))))),
});
return name;
}
/// Builds the polyfill function for the `atanh` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol atanh(const type::Type* ty) {
auto name = b.Symbols().New("tint_atanh");
uint32_t width = WidthOf(ty);
auto V = [&](AFloat value) -> const ast::Expression* {
const ast::Expression* expr = b.Expr(value);
if (width == 1) {
return expr;
}
return b.Call(T(ty), expr);
};
utils::Vector<const ast::Statement*, 1> body;
switch (cfg.builtins.atanh) {
case Level::kFull:
// return log((1+x) / (1-x)) * 0.5
body.Push(
b.Return(b.Mul(b.Call("log", b.Div(b.Add(1_a, "x"), b.Sub(1_a, "x"))), 0.5_a)));
break;
case Level::kRangeCheck:
// return select(atanh(x), 0, x >= 1);
body.Push(b.Return(b.Call("select", b.Call("atanh", "x"), V(0.0_a),
b.GreaterThanEqual("x", V(1.0_a)))));
break;
default:
TINT_ICE(Transform, b.Diagnostics())
<< "unhandled polyfill level: " << static_cast<int>(cfg.builtins.acosh);
return {};
}
b.Func(name, utils::Vector{b.Param("x", T(ty))}, T(ty), body);
return name;
}
/// Builds the polyfill function for the `clamp` builtin when called with integer arguments
/// (scalar or vector)
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol clampInteger(const type::Type* ty) {
auto name = b.Symbols().New("tint_clamp");
b.Func(name,
utils::Vector{
b.Param("e", T(ty)),
b.Param("low", T(ty)),
b.Param("high", T(ty)),
},
T(ty),
utils::Vector{
// return min(max(e, low), high);
b.Return(b.Call("min", b.Call("max", "e", "low"), "high")),
});
return name;
}
/// Builds the polyfill function for the `countLeadingZeros` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol countLeadingZeros(const type::Type* ty) {
auto name = b.Symbols().New("tint_count_leading_zeros");
uint32_t width = WidthOf(ty);
// Returns either u32 or vecN<u32>
auto U = [&]() {
if (width == 1) {
return b.ty.u32();
}
return b.ty.vec<u32>(width);
};
auto V = [&](uint32_t value) -> const ast::Expression* {
return ScalarOrVector(width, u32(value));
};
b.Func(
name,
utils::Vector{
b.Param("v", T(ty)),
},
T(ty),
utils::Vector{
// var x = U(v);
b.Decl(b.Var("x", b.Call(U(), b.Expr("v")))),
// let b16 = select(0, 16, x <= 0x0000ffff);
b.Decl(b.Let("b16",
b.Call("select", V(0), V(16), b.LessThanEqual("x", V(0x0000ffff))))),
// x = x << b16;
b.Assign("x", b.Shl("x", "b16")),
// let b8 = select(0, 8, x <= 0x00ffffff);
b.Decl(
b.Let("b8", b.Call("select", V(0), V(8), b.LessThanEqual("x", V(0x00ffffff))))),
// x = x << b8;
b.Assign("x", b.Shl("x", "b8")),
// let b4 = select(0, 4, x <= 0x0fffffff);
b.Decl(
b.Let("b4", b.Call("select", V(0), V(4), b.LessThanEqual("x", V(0x0fffffff))))),
// x = x << b4;
b.Assign("x", b.Shl("x", "b4")),
// let b2 = select(0, 2, x <= 0x3fffffff);
b.Decl(
b.Let("b2", b.Call("select", V(0), V(2), b.LessThanEqual("x", V(0x3fffffff))))),
// x = x << b2;
b.Assign("x", b.Shl("x", "b2")),
// let b1 = select(0, 1, x <= 0x7fffffff);
b.Decl(
b.Let("b1", b.Call("select", V(0), V(1), b.LessThanEqual("x", V(0x7fffffff))))),
// let is_zero = select(0, 1, x == 0);
b.Decl(b.Let("is_zero", b.Call("select", V(0), V(1), b.Equal("x", V(0))))),
// return R((b16 | b8 | b4 | b2 | b1) + zero);
b.Return(b.Call(T(ty), b.Add(b.Or(b.Or(b.Or(b.Or("b16", "b8"), "b4"), "b2"), "b1"),
"is_zero"))),
});
return name;
}
/// Builds the polyfill function for the `countTrailingZeros` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol countTrailingZeros(const type::Type* ty) {
auto name = b.Symbols().New("tint_count_trailing_zeros");
uint32_t width = WidthOf(ty);
// Returns either u32 or vecN<u32>
auto U = [&]() {
if (width == 1) {
return b.ty.u32();
}
return b.ty.vec<u32>(width);
};
auto V = [&](uint32_t value) -> const ast::Expression* {
return ScalarOrVector(width, u32(value));
};
auto B = [&](const ast::Expression* value) -> const ast::Expression* {
if (width == 1) {
return b.Call<bool>(value);
}
return b.Call(b.ty.vec<bool>(width), value);
};
b.Func(
name,
utils::Vector{
b.Param("v", T(ty)),
},
T(ty),
utils::Vector{
// var x = U(v);
b.Decl(b.Var("x", b.Call(U(), b.Expr("v")))),
// let b16 = select(16, 0, bool(x & 0x0000ffff));
b.Decl(b.Let("b16", b.Call("select", V(16), V(0), B(b.And("x", V(0x0000ffff)))))),
// x = x >> b16;
b.Assign("x", b.Shr("x", "b16")),
// let b8 = select(8, 0, bool(x & 0x000000ff));
b.Decl(b.Let("b8", b.Call("select", V(8), V(0), B(b.And("x", V(0x000000ff)))))),
// x = x >> b8;
b.Assign("x", b.Shr("x", "b8")),
// let b4 = select(4, 0, bool(x & 0x0000000f));
b.Decl(b.Let("b4", b.Call("select", V(4), V(0), B(b.And("x", V(0x0000000f)))))),
// x = x >> b4;
b.Assign("x", b.Shr("x", "b4")),
// let b2 = select(2, 0, bool(x & 0x00000003));
b.Decl(b.Let("b2", b.Call("select", V(2), V(0), B(b.And("x", V(0x00000003)))))),
// x = x >> b2;
b.Assign("x", b.Shr("x", "b2")),
// let b1 = select(1, 0, bool(x & 0x00000001));
b.Decl(b.Let("b1", b.Call("select", V(1), V(0), B(b.And("x", V(0x00000001)))))),
// let is_zero = select(0, 1, x == 0);
b.Decl(b.Let("is_zero", b.Call("select", V(0), V(1), b.Equal("x", V(0))))),
// return R((b16 | b8 | b4 | b2 | b1) + zero);
b.Return(b.Call(T(ty), b.Add(b.Or(b.Or(b.Or(b.Or("b16", "b8"), "b4"), "b2"), "b1"),
"is_zero"))),
});
return name;
}
/// Builds the polyfill function for the `extractBits` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol extractBits(const type::Type* ty) {
auto name = b.Symbols().New("tint_extract_bits");
uint32_t width = WidthOf(ty);
constexpr uint32_t W = 32u; // 32-bit
auto vecN_u32 = [&](const ast::Expression* value) -> const ast::Expression* {
if (width == 1) {
return value;
}
return b.Call(b.ty.vec<u32>(width), value);
};
utils::Vector<const ast::Statement*, 8> body{
b.Decl(b.Let("s", b.Call("min", "offset", u32(W)))),
b.Decl(b.Let("e", b.Call("min", u32(W), b.Add("s", "count")))),
};
switch (cfg.builtins.extract_bits) {
case Level::kFull:
body.Push(b.Decl(b.Let("shl", b.Sub(u32(W), "e"))));
body.Push(b.Decl(b.Let("shr", b.Add("shl", "s"))));
// Here we don't want the shl and shr modulos the rhs, so handle the `rhs >= 32u`
// cases using `select`. In order to handle the signed shr `lhs >> rhs` corrently,
// use `(lhs >> 31u) >> 1u` if `rhs >= 32u`.
body.Push(b.Decl(b.Let("shl_result", b.Call("select", b.Call(T(ty)),
b.Shl("v", vecN_u32(b.Expr("shl"))),
b.LessThan("shl", 32_u)))));
body.Push(b.Return(b.Call(
"select",
b.Shr(b.Shr("shl_result", vecN_u32(b.Expr(31_u))), vecN_u32(b.Expr(1_u))),
b.Shr("shl_result", vecN_u32(b.Expr("shr"))), b.LessThan("shr", 32_u))
));
break;
case Level::kClampParameters:
body.Push(b.Return(b.Call("extractBits", "v", "s", b.Sub("e", "s"))));
break;
default:
TINT_ICE(Transform, b.Diagnostics())
<< "unhandled polyfill level: " << static_cast<int>(cfg.builtins.extract_bits);
return {};
}
b.Func(name,
utils::Vector{
b.Param("v", T(ty)),
b.Param("offset", b.ty.u32()),
b.Param("count", b.ty.u32()),
},
T(ty), std::move(body));
return name;
}
/// Builds the polyfill function for the `firstLeadingBit` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol firstLeadingBit(const type::Type* ty) {
auto name = b.Symbols().New("tint_first_leading_bit");
uint32_t width = WidthOf(ty);
// Returns either u32 or vecN<u32>
auto U = [&]() {
if (width == 1) {
return b.ty.u32();
}
return b.ty.vec<u32>(width);
};
auto V = [&](uint32_t value) -> const ast::Expression* {
return ScalarOrVector(width, u32(value));
};
auto B = [&](const ast::Expression* value) -> const ast::Expression* {
if (width == 1) {
return b.Call<bool>(value);
}
return b.Call(b.ty.vec<bool>(width), value);
};
const ast::Expression* x = nullptr;
if (ty->is_unsigned_integer_scalar_or_vector()) {
x = b.Expr("v");
} else {
// If ty is signed, then the value is inverted if the sign is negative
x = b.Call("select", //
b.Call(U(), "v"), //
b.Call(U(), b.Complement("v")), //
b.LessThan("v", ScalarOrVector(width, 0_i)));
}
b.Func(
name,
utils::Vector{
b.Param("v", T(ty)),
},
T(ty),
utils::Vector{
// var x = v; (unsigned)
// var x = select(U(v), ~U(v), v < 0); (signed)
b.Decl(b.Var("x", x)),
// let b16 = select(0, 16, bool(x & 0xffff0000));
b.Decl(b.Let("b16", b.Call("select", V(0), V(16), B(b.And("x", V(0xffff0000)))))),
// x = x >> b16;
b.Assign("x", b.Shr("x", "b16")),
// let b8 = select(0, 8, bool(x & 0x0000ff00));
b.Decl(b.Let("b8", b.Call("select", V(0), V(8), B(b.And("x", V(0x0000ff00)))))),
// x = x >> b8;
b.Assign("x", b.Shr("x", "b8")),
// let b4 = select(0, 4, bool(x & 0x000000f0));
b.Decl(b.Let("b4", b.Call("select", V(0), V(4), B(b.And("x", V(0x000000f0)))))),
// x = x >> b4;
b.Assign("x", b.Shr("x", "b4")),
// let b2 = select(0, 2, bool(x & 0x0000000c));
b.Decl(b.Let("b2", b.Call("select", V(0), V(2), B(b.And("x", V(0x0000000c)))))),
// x = x >> b2;
b.Assign("x", b.Shr("x", "b2")),
// let b1 = select(0, 1, bool(x & 0x00000002));
b.Decl(b.Let("b1", b.Call("select", V(0), V(1), B(b.And("x", V(0x00000002)))))),
// let is_zero = select(0, 0xffffffff, x == 0);
b.Decl(b.Let("is_zero", b.Call("select", V(0), V(0xffffffff), b.Equal("x", V(0))))),
// return R(b16 | b8 | b4 | b2 | b1 | zero);
b.Return(b.Call(
T(ty), b.Or(b.Or(b.Or(b.Or(b.Or("b16", "b8"), "b4"), "b2"), "b1"), "is_zero"))),
});
return name;
}
/// Builds the polyfill function for the `firstTrailingBit` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol firstTrailingBit(const type::Type* ty) {
auto name = b.Symbols().New("tint_first_trailing_bit");
uint32_t width = WidthOf(ty);
// Returns either u32 or vecN<u32>
auto U = [&]() {
if (width == 1) {
return b.ty.u32();
}
return b.ty.vec<u32>(width);
};
auto V = [&](uint32_t value) -> const ast::Expression* {
return ScalarOrVector(width, u32(value));
};
auto B = [&](const ast::Expression* value) -> const ast::Expression* {
if (width == 1) {
return b.Call<bool>(value);
}
return b.Call(b.ty.vec<bool>(width), value);
};
b.Func(
name,
utils::Vector{
b.Param("v", T(ty)),
},
T(ty),
utils::Vector{
// var x = U(v);
b.Decl(b.Var("x", b.Call(U(), b.Expr("v")))),
// let b16 = select(16, 0, bool(x & 0x0000ffff));
b.Decl(b.Let("b16", b.Call("select", V(16), V(0), B(b.And("x", V(0x0000ffff)))))),
// x = x >> b16;
b.Assign("x", b.Shr("x", "b16")),
// let b8 = select(8, 0, bool(x & 0x000000ff));
b.Decl(b.Let("b8", b.Call("select", V(8), V(0), B(b.And("x", V(0x000000ff)))))),
// x = x >> b8;
b.Assign("x", b.Shr("x", "b8")),
// let b4 = select(4, 0, bool(x & 0x0000000f));
b.Decl(b.Let("b4", b.Call("select", V(4), V(0), B(b.And("x", V(0x0000000f)))))),
// x = x >> b4;
b.Assign("x", b.Shr("x", "b4")),
// let b2 = select(2, 0, bool(x & 0x00000003));
b.Decl(b.Let("b2", b.Call("select", V(2), V(0), B(b.And("x", V(0x00000003)))))),
// x = x >> b2;
b.Assign("x", b.Shr("x", "b2")),
// let b1 = select(1, 0, bool(x & 0x00000001));
b.Decl(b.Let("b1", b.Call("select", V(1), V(0), B(b.And("x", V(0x00000001)))))),
// let is_zero = select(0, 0xffffffff, x == 0);
b.Decl(b.Let("is_zero", b.Call("select", V(0), V(0xffffffff), b.Equal("x", V(0))))),
// return R(b16 | b8 | b4 | b2 | b1 | is_zero);
b.Return(b.Call(
T(ty), b.Or(b.Or(b.Or(b.Or(b.Or("b16", "b8"), "b4"), "b2"), "b1"), "is_zero"))),
});
return name;
}
/// Builds the polyfill function for the `insertBits` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol insertBits(const type::Type* ty) {
auto name = b.Symbols().New("tint_insert_bits");
uint32_t width = WidthOf(ty);
// Currently in WGSL parameters of insertBits must be i32, u32, vecN<i32> or vecN<u32>
if (TINT_UNLIKELY(((!type::Type::DeepestElementOf(ty)->IsAnyOf<type::I32, type::U32>())))) {
TINT_ICE(Transform, b.Diagnostics())
<< "insertBits polyfill only support i32, u32, and vector of i32 or u32, got "
<< b.FriendlyName(ty);
return {};
}
constexpr uint32_t W = 32u; // 32-bit
auto V = [&](auto value) -> const ast::Expression* {
const ast::Expression* expr = b.Expr(value);
if (!ty->is_unsigned_integer_scalar_or_vector()) {
expr = b.Call<i32>(expr);
}
if (ty->Is<type::Vector>()) {
expr = b.Call(T(ty), expr);
}
return expr;
};
auto U = [&](auto value) -> const ast::Expression* {
if (width == 1) {
return b.Expr(value);
}
return b.vec(b.ty.u32(), width, value);
};
// Polyfill algorithm:
// s = min(offset, 32u);
// e = min(32u, (s + count));
// mask = (((1u << s) - 1u) ^ ((1u << e) - 1u));
// return (((n << s) & mask) | (v & ~(mask)));
// Note that the algorithm above use the left-shifting in C++ manner, but in WGSL, HLSL, MSL
// the rhs are modulo to bit-width of lhs (that is 32u in this case), and in GLSL the result
// is undefined if rhs is greater than or equal to bit-width of lhs. The results of `x << y`
// in C++ and HLSL are different when `y >= 32u`, and the `s` and `e` defined above can be
// 32u, which are cases we must handle specially. Replace all `(x << y)` to
// `select(Tx(), x << y, y < 32u)`, in which `Tx` is the type of x, where y can be greater
// than or equal to 32u.
// WGSL polyfill function:
// fn tint_insert_bits(v : T, n : T, offset : u32, count : u32) -> T {
// let e = offset + count;
// let mask = (
// (select(0u, 1u << offset, offset < 32u) - 1u) ^
// (select(0u, 1u << e, e < 32u) - 1u)
// );
// return ((select(T(), n << offset, offset < 32u) & mask) | (v & ~(mask)));
// }
utils::Vector<const ast::Statement*, 8> body;
switch (cfg.builtins.insert_bits) {
case Level::kFull:
// let e = offset + count;
body.Push(b.Decl(b.Let("e", b.Add("offset", "count"))));
// let mask = (
// (select(0u, 1u << offset, offset < 32u) - 1u) ^
// (select(0u, 1u << e, e < 32u) - 1u)
// );
body.Push(b.Decl(b.Let(
"mask",
b.Xor( //
b.Sub(
b.Call("select", 0_u, b.Shl(1_u, "offset"), b.LessThan("offset", 32_u)),
1_u),
b.Sub(b.Call("select", 0_u, b.Shl(1_u, "e"), b.LessThan("e", 32_u)),
1_u) //
))));
// return ((select(T(), n << offset, offset < 32u) & mask) | (v & ~(mask)));
body.Push(
b.Return(b.Or(b.And(b.Call("select", b.Call(T(ty)), b.Shl("n", U("offset")),
b.LessThan("offset", 32_u)),
V("mask")),
b.And("v", V(b.Complement("mask"))))));
break;
case Level::kClampParameters:
body.Push(b.Decl(b.Let("s", b.Call("min", "offset", u32(W)))));
body.Push(b.Decl(b.Let("e", b.Call("min", u32(W), b.Add("s", "count")))));
body.Push(b.Return(b.Call("insertBits", "v", "n", "s", b.Sub("e", "s"))));
break;
default:
TINT_ICE(Transform, b.Diagnostics())
<< "unhandled polyfill level: " << static_cast<int>(cfg.builtins.insert_bits);
return {};
}
b.Func(name,
utils::Vector{
b.Param("v", T(ty)),
b.Param("n", T(ty)),
b.Param("offset", b.ty.u32()),
b.Param("count", b.ty.u32()),
},
T(ty), body);
return name;
}
/// Builds the polyfill function for the `reflect` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol reflect(const type::Type* ty) {
auto name = b.Symbols().New("tint_reflect");
// WGSL polyfill function:
// fn tint_reflect(e1 : T, e2 : T) -> T {
// let factor = (-2.0 * dot(e1, e2));
// return (e1 + (factor * e2));
// }
// Using -2.0 instead of 2.0 in factor to prevent the optimization that cause wrong result.
// See https://crbug.com/tint/1798 for more details.
auto body = utils::Vector{
b.Decl(b.Let("factor", b.Mul(-2.0_a, b.Call("dot", "e1", "e2")))),
b.Return(b.Add("e1", b.Mul("factor", "e2"))),
};
b.Func(name,
utils::Vector{
b.Param("e1", T(ty)),
b.Param("e2", T(ty)),
},
T(ty), body);
return name;
}
/// Builds the polyfill function for the `saturate` builtin
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol saturate(const type::Type* ty) {
auto name = b.Symbols().New("tint_saturate");
auto body = utils::Vector{
b.Return(b.Call("clamp", "v", b.Call(T(ty), 0_a), b.Call(T(ty), 1_a))),
};
b.Func(name,
utils::Vector{
b.Param("v", T(ty)),
},
T(ty), body);
return name;
}
/// Builds the polyfill function for the `sign` builtin when the element type is integer
/// @param ty the parameter and return type for the function
/// @return the polyfill function name
Symbol sign_int(const type::Type* ty) {
const uint32_t width = WidthOf(ty);
auto zero = [&] { return ScalarOrVector(width, 0_a); };
// pos_or_neg_one = (v > 0) ? 1 : -1
auto pos_or_neg_one = b.Call("select", //
ScalarOrVector(width, -1_a), //
ScalarOrVector(width, 1_a), //
b.GreaterThan("v", zero()));
auto name = b.Symbols().New("tint_sign");
b.Func(name,
utils::Vector{
b.Param("v", T(ty)),
},
T(ty),
utils::Vector{
b.Return(b.Call("select", pos_or_neg_one, zero(), b.Equal("v", zero()))),
});
return name;
}
/// Builds the polyfill function for the `textureSampleBaseClampToEdge` builtin, when the
/// texture type is texture_2d<f32>.
/// @return the polyfill function name
Symbol textureSampleBaseClampToEdge_2d_f32() {
auto name = b.Symbols().New("tint_textureSampleBaseClampToEdge");
auto body = utils::Vector{
b.Decl(b.Let("dims", b.Call(b.ty.vec2<f32>(), b.Call("textureDimensions", "t", 0_a)))),
b.Decl(b.Let("half_texel", b.Div(b.vec2<f32>(0.5_a), "dims"))),
b.Decl(
b.Let("clamped", b.Call("clamp", "coord", "half_texel", b.Sub(1_a, "half_texel")))),
b.Return(b.Call("textureSampleLevel", "t", "s", "clamped", 0_a)),
};
b.Func(name,
utils::Vector{
b.Param("t", b.ty.sampled_texture(type::TextureDimension::k2d, b.ty.f32())),
b.Param("s", b.ty.sampler(type::SamplerKind::kSampler)),
b.Param("coord", b.ty.vec2<f32>()),
},
b.ty.vec4<f32>(), body);
return name;
}
/// Builds the polyfill function for the `quantizeToF16` builtin, by replacing the vector form
/// with scalar calls.
/// @param vec the vector type
/// @return the polyfill function name
Symbol quantizeToF16(const type::Vector* vec) {
auto name = b.Symbols().New("tint_quantizeToF16");
utils::Vector<const ast::Expression*, 4> args;
for (uint32_t i = 0; i < vec->Width(); i++) {
args.Push(b.Call("quantizeToF16", b.IndexAccessor("v", u32(i))));
}
b.Func(name,
utils::Vector{
b.Param("v", T(vec)),
},
T(vec),
utils::Vector{
b.Return(b.Call(T(vec), std::move(args))),
});
return name;
}
/// Builds the polyfill function for the `workgroupUniformLoad` builtin.
/// @param type the type being loaded
/// @return the polyfill function name
Symbol workgroupUniformLoad(const type::Type* type) {
if (!has_full_ptr_params) {
b.Enable(builtin::Extension::kChromiumExperimentalFullPtrParameters);
has_full_ptr_params = true;
}
auto name = b.Symbols().New("tint_workgroupUniformLoad");
b.Func(name,
utils::Vector{
b.Param("p", b.ty.pointer(T(type), builtin::AddressSpace::kWorkgroup)),
},
T(type),
utils::Vector{
b.CallStmt(b.Call("workgroupBarrier")),
b.Decl(b.Let("result", b.Deref("p"))),
b.CallStmt(b.Call("workgroupBarrier")),
b.Return("result"),
});
return name;
}
/// Builds the polyfill function to value convert a scalar or vector of f32 to an i32 or u32 (or
/// vector of).
/// @param source the type of the value being converted
/// @param target the target conversion type
/// @return the polyfill function name
Symbol ConvF32ToIU32(const type::Type* source, const type::Type* target) {
struct Limits {
AFloat low_condition;
AInt low_limit;
AFloat high_condition;
AInt high_limit;
};
const bool is_signed = target->is_signed_integer_scalar_or_vector();
const Limits limits = is_signed ? Limits{
/* low_condition */ -AFloat(0x80000000),
/* low_limit */ -AInt(0x80000000),
/* high_condition */ AFloat(0x7fffff80),
/* high_limit */ AInt(0x7fffffff),
}
: Limits{
/* low_condition */ AFloat(0),
/* low_limit */ AInt(0),
/* high_condition */ AFloat(0xffffff00),
/* high_limit */ AInt(0xffffffff),
};
const uint32_t width = WidthOf(target);
// select(target(v), low_limit, v < low_condition)
auto* select_low = b.Call(builtin::Function::kSelect, //
b.Call(T(target), "v"), //
ScalarOrVector(width, limits.low_limit), //
b.LessThan("v", ScalarOrVector(width, limits.low_condition)));
// select(high_limit, select_low, v < high_condition)
auto* select_high = b.Call(builtin::Function::kSelect, //
ScalarOrVector(width, limits.high_limit), //
select_low, //
b.LessThan("v", ScalarOrVector(width, limits.high_condition)));
auto name = b.Symbols().New(is_signed ? "tint_ftoi" : "tint_ftou");
b.Func(name, utils::Vector{b.Param("v", T(source))}, T(target),
utils::Vector{b.Return(select_high)});
return name;
}
////////////////////////////////////////////////////////////////////////////
// Inline polyfills
////////////////////////////////////////////////////////////////////////////
/// Builds the polyfill inline expression for a bitshift left or bitshift right, ensuring that
/// the RHS is modulo the bit-width of the LHS.
/// @param bin_op the original BinaryExpression
/// @return the polyfill value for bitshift operation
const ast::Expression* BitshiftModulo(const ast::BinaryExpression* bin_op) {
auto* lhs_ty = src->TypeOf(bin_op->lhs)->UnwrapRef();
auto* rhs_ty = src->TypeOf(bin_op->rhs)->UnwrapRef();
auto* lhs_el_ty = type::Type::DeepestElementOf(lhs_ty);
const ast::Expression* mask = b.Expr(AInt(lhs_el_ty->Size() * 8 - 1));
if (rhs_ty->Is<type::Vector>()) {
mask = b.Call(CreateASTTypeFor(ctx, rhs_ty), mask);
}
auto* lhs = ctx.Clone(bin_op->lhs);
auto* rhs = b.And(ctx.Clone(bin_op->rhs), mask);
return b.create<ast::BinaryExpression>(ctx.Clone(bin_op->source), bin_op->op, lhs, rhs);
}
/// Builds the polyfill inline expression for a integer divide or modulo, preventing DBZs and
/// integer overflows.
/// @param bin_op the original BinaryExpression
/// @return the polyfill divide or modulo
const ast::Expression* IntDivMod(const ast::BinaryExpression* bin_op) {
auto* lhs_ty = src->TypeOf(bin_op->lhs)->UnwrapRef();
auto* rhs_ty = src->TypeOf(bin_op->rhs)->UnwrapRef();
BinaryOpSignature sig{bin_op->op, lhs_ty, rhs_ty};
auto fn = binary_op_polyfills.GetOrCreate(sig, [&] {
const bool is_div = bin_op->op == ast::BinaryOp::kDivide;
uint32_t lhs_width = 1;
uint32_t rhs_width = 1;
const auto* lhs_el_ty = type::Type::ElementOf(lhs_ty, &lhs_width);
const auto* rhs_el_ty = type::Type::ElementOf(rhs_ty, &rhs_width);
const uint32_t width = std::max(lhs_width, rhs_width);
const char* lhs = "lhs";
const char* rhs = "rhs";
utils::Vector<const ast::Statement*, 4> body;
if (lhs_width < width) {
// lhs is scalar, rhs is vector. Convert lhs to vector.
body.Push(b.Decl(b.Let("l", b.vec(T(lhs_el_ty), width, b.Expr(lhs)))));
lhs = "l";
}
if (rhs_width < width) {
// lhs is vector, rhs is scalar. Convert rhs to vector.
body.Push(b.Decl(b.Let("r", b.vec(T(rhs_el_ty), width, b.Expr(rhs)))));
rhs = "r";
}
auto name = b.Symbols().New(is_div ? "tint_div" : "tint_mod");
auto* rhs_is_zero = b.Equal(rhs, ScalarOrVector(width, 0_a));
if (lhs_ty->is_signed_integer_scalar_or_vector()) {
const auto bits = lhs_el_ty->Size() * 8;
auto min_int = AInt(AInt::kLowestValue >> (AInt::kNumBits - bits));
const ast::Expression* lhs_is_min = b.Equal(lhs, ScalarOrVector(width, min_int));
const ast::Expression* rhs_is_minus_one = b.Equal(rhs, ScalarOrVector(width, -1_a));
// use_one = rhs_is_zero | ((lhs == MIN_INT) & (rhs == -1))
auto* use_one = b.Or(rhs_is_zero, b.And(lhs_is_min, rhs_is_minus_one));
// Special handling for mod in case either operand is negative, as negative operands
// for % is undefined behaviour for most backends (HLSL, MSL, GLSL, SPIR-V).
if (!is_div) {
const char* rhs_or_one = "rhs_or_one";
body.Push(b.Decl(b.Let(
rhs_or_one, b.Call("select", rhs, ScalarOrVector(width, 1_a), use_one))));
// Is either operand negative?
// (lhs | rhs) & (1<<31)
auto sign_bit_mask = ScalarOrVector(width, u32(1 << (bits - 1)));
auto* lhs_or_rhs = CastScalarOrVector<u32>(width, b.Or(lhs, rhs_or_one));
auto* lhs_or_rhs_is_neg =
b.NotEqual(b.And(lhs_or_rhs, sign_bit_mask), ScalarOrVector(width, 0_u));
// lhs - trunc(lhs / rhs) * rhs (note: integral division truncates)
auto* slow_mod = b.Sub(lhs, b.Mul(b.Div(lhs, rhs_or_one), rhs_or_one));
// lhs % rhs
auto* fast_mod = b.Mod(lhs, rhs_or_one);
auto* use_slow = b.Call("any", lhs_or_rhs_is_neg);
body.Push(b.If(use_slow, b.Block(b.Return(slow_mod)),
b.Else(b.Block(b.Return(fast_mod)))));
} else {
auto* rhs_or_one = b.Call("select", rhs, ScalarOrVector(width, 1_a), use_one);
body.Push(b.Return(is_div ? b.Div(lhs, rhs_or_one) : b.Mod(lhs, rhs_or_one)));
}
} else {
auto* rhs_or_one = b.Call("select", rhs, ScalarOrVector(width, 1_a), rhs_is_zero);
body.Push(b.Return(is_div ? b.Div(lhs, rhs_or_one) : b.Mod(lhs, rhs_or_one)));
}
b.Func(name,
utils::Vector{
b.Param("lhs", T(lhs_ty)),
b.Param("rhs", T(rhs_ty)),
},
width == 1 ? T(lhs_ty) : b.ty.vec(T(lhs_el_ty), width), // return type
std::move(body));
return name;
});
auto* lhs = ctx.Clone(bin_op->lhs);
auto* rhs = ctx.Clone(bin_op->rhs);
return b.Call(fn, lhs, rhs);
}
/// Builds the polyfill inline expression for a precise float modulo, as defined in the spec.
/// @param bin_op the original BinaryExpression
/// @return the polyfill divide or modulo
const ast::Expression* PreciseFloatMod(const ast::BinaryExpression* bin_op) {
auto* lhs_ty = src->TypeOf(bin_op->lhs)->UnwrapRef();
auto* rhs_ty = src->TypeOf(bin_op->rhs)->UnwrapRef();
BinaryOpSignature sig{bin_op->op, lhs_ty, rhs_ty};
auto fn = binary_op_polyfills.GetOrCreate(sig, [&] {
uint32_t lhs_width = 1;
uint32_t rhs_width = 1;
const auto* lhs_el_ty = type::Type::ElementOf(lhs_ty, &lhs_width);
const auto* rhs_el_ty = type::Type::ElementOf(rhs_ty, &rhs_width);
const uint32_t width = std::max(lhs_width, rhs_width);
const char* lhs = "lhs";
const char* rhs = "rhs";
utils::Vector<const ast::Statement*, 4> body;
if (lhs_width < width) {
// lhs is scalar, rhs is vector. Convert lhs to vector.
body.Push(b.Decl(b.Let("l", b.vec(T(lhs_el_ty), width, b.Expr(lhs)))));
lhs = "l";
}
if (rhs_width < width) {
// lhs is vector, rhs is scalar. Convert rhs to vector.
body.Push(b.Decl(b.Let("r", b.vec(T(rhs_el_ty), width, b.Expr(rhs)))));
rhs = "r";
}
auto name = b.Symbols().New("tint_float_mod");
// lhs - trunc(lhs / rhs) * rhs
auto* precise_mod = b.Sub(lhs, b.Mul(b.Call("trunc", b.Div(lhs, rhs)), rhs));
body.Push(b.Return(precise_mod));
b.Func(name,
utils::Vector{
b.Param("lhs", T(lhs_ty)),
b.Param("rhs", T(rhs_ty)),
},
width == 1 ? T(lhs_ty) : b.ty.vec(T(lhs_el_ty), width), // return type
std::move(body));
return name;
});
auto* lhs = ctx.Clone(bin_op->lhs);
auto* rhs = ctx.Clone(bin_op->rhs);
return b.Call(fn, lhs, rhs);
}
/// @returns the AST type for the given sem type
ast::Type T(const type::Type* ty) { return CreateASTTypeFor(ctx, ty); }
/// @returns 1 if `ty` is not a vector, otherwise the vector width
uint32_t WidthOf(const type::Type* ty) const {
if (auto* v = ty->As<type::Vector>()) {
return v->Width();
}
return 1;
}
/// @returns a scalar or vector with the given width, with each element with
/// the given value.
template <typename T>
const ast::Expression* ScalarOrVector(uint32_t width, T value) {
if (width == 1) {
return b.Expr(value);
}
return b.Call(b.ty.vec<T>(width), value);
}
template <typename To>
const ast::Expression* CastScalarOrVector(uint32_t width, const ast::Expression* e) {
if (width == 1) {
return b.Call(b.ty.Of<To>(), e);
}
return b.Call(b.ty.vec<To>(width), e);
}
/// Examines the call expression @p expr, applying any necessary polyfill transforms
void Call(const ast::CallExpression* expr) {
auto* call = src->Sem().Get(expr)->UnwrapMaterialize()->As<sem::Call>();
if (!call || call->Stage() == sem::EvaluationStage::kConstant ||
call->Stage() == sem::EvaluationStage::kNotEvaluated) {
return; // Don't polyfill @const expressions
}
Symbol fn = Switch(
call->Target(), //
[&](const sem::Builtin* builtin) {
switch (builtin->Type()) {
case builtin::Function::kAcosh:
if (cfg.builtins.acosh != Level::kNone) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return acosh(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kAsinh:
if (cfg.builtins.asinh) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return asinh(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kAtanh:
if (cfg.builtins.atanh != Level::kNone) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return atanh(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kClamp:
if (cfg.builtins.clamp_int) {
auto& sig = builtin->Signature();
if (sig.parameters[0]->Type()->is_integer_scalar_or_vector()) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return clampInteger(builtin->ReturnType()); });
}
}
return Symbol{};
case builtin::Function::kCountLeadingZeros:
if (cfg.builtins.count_leading_zeros) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return countLeadingZeros(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kCountTrailingZeros:
if (cfg.builtins.count_trailing_zeros) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return countTrailingZeros(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kExtractBits:
if (cfg.builtins.extract_bits != Level::kNone) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return extractBits(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kFirstLeadingBit:
if (cfg.builtins.first_leading_bit) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return firstLeadingBit(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kFirstTrailingBit:
if (cfg.builtins.first_trailing_bit) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return firstTrailingBit(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kInsertBits:
if (cfg.builtins.insert_bits != Level::kNone) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return insertBits(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kReflect:
// Only polyfill for vec2<f32>. See https://crbug.com/tint/1798 for
// more details.
if (cfg.builtins.reflect_vec2_f32) {
auto& sig = builtin->Signature();
auto* vec = sig.return_type->As<type::Vector>();
if (vec && vec->Width() == 2 && vec->type()->Is<type::F32>()) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return reflect(builtin->ReturnType()); });
}
}
return Symbol{};
case builtin::Function::kSaturate:
if (cfg.builtins.saturate) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return saturate(builtin->ReturnType()); });
}
return Symbol{};
case builtin::Function::kSign:
if (cfg.builtins.sign_int) {
auto* ty = builtin->ReturnType();
if (ty->is_signed_integer_scalar_or_vector()) {
return builtin_polyfills.GetOrCreate(builtin,
[&] { return sign_int(ty); });
}
}
return Symbol{};
case builtin::Function::kTextureSampleBaseClampToEdge:
if (cfg.builtins.texture_sample_base_clamp_to_edge_2d_f32) {
auto& sig = builtin->Signature();
auto* tex = sig.Parameter(sem::ParameterUsage::kTexture);
if (auto* stex = tex->Type()->As<type::SampledTexture>()) {
if (stex->type()->Is<type::F32>()) {
return builtin_polyfills.GetOrCreate(builtin, [&] {
return textureSampleBaseClampToEdge_2d_f32();
});
}
}
}
return Symbol{};
case builtin::Function::kTextureStore:
if (cfg.builtins.bgra8unorm) {
auto& sig = builtin->Signature();
auto* tex = sig.Parameter(sem::ParameterUsage::kTexture);
if (auto* stex = tex->Type()->As<type::StorageTexture>()) {
if (stex->texel_format() == builtin::TexelFormat::kBgra8Unorm) {
size_t value_idx = static_cast<size_t>(
sig.IndexOf(sem::ParameterUsage::kValue));
ctx.Replace(expr, [this, expr, value_idx] {
utils::Vector<const ast::Expression*, 3> args;
for (auto* arg : expr->args) {
arg = ctx.Clone(arg);
if (args.Length() == value_idx) { // value
arg = ctx.dst->MemberAccessor(arg, "bgra");
}
args.Push(arg);
}
return ctx.dst->Call(
utils::ToString(builtin::Function::kTextureStore),
std::move(args));
});
made_changes = true;
}
}
}
return Symbol{};
case builtin::Function::kQuantizeToF16:
if (cfg.builtins.quantize_to_vec_f16) {
if (auto* vec = builtin->ReturnType()->As<type::Vector>()) {
return builtin_polyfills.GetOrCreate(
builtin, [&] { return quantizeToF16(vec); });
}
}
return Symbol{};
case builtin::Function::kWorkgroupUniformLoad:
if (cfg.builtins.workgroup_uniform_load) {
return builtin_polyfills.GetOrCreate(builtin, [&] {
return workgroupUniformLoad(builtin->ReturnType());
});
}
return Symbol{};
default:
return Symbol{};
}
},
[&](const sem::ValueConversion* conv) {
if (cfg.builtins.conv_f32_to_iu32) {
auto* src_ty = conv->Source();
if (tint::Is<type::F32>(type::Type::ElementOf(src_ty))) {
auto* dst_ty = conv->Target();
if (tint::IsAnyOf<type::I32, type::U32>(type::Type::ElementOf(dst_ty))) {
return f32_conv_polyfills.GetOrCreate(dst_ty, [&] { //
return ConvF32ToIU32(src_ty, dst_ty);
});
}
}
}
return Symbol{};
});
if (fn.IsValid()) {
ctx.Replace(call->Declaration(),
[this, fn, expr] { return ctx.dst->Call(fn, ctx.Clone(expr->args)); });
made_changes = true;
}
}
};
BuiltinPolyfill::BuiltinPolyfill() = default;
BuiltinPolyfill::~BuiltinPolyfill() = default;
Transform::ApplyResult BuiltinPolyfill::Apply(const Program* src,
const DataMap& data,
DataMap&) const {
auto* cfg = data.Get<Config>();
if (!cfg) {
return SkipTransform;
}
return State{src, *cfg}.Run();
}
BuiltinPolyfill::Config::Config(const Builtins& b) : builtins(b) {}
BuiltinPolyfill::Config::Config(const Config&) = default;
BuiltinPolyfill::Config::~Config() = default;
} // namespace tint::transform